#2768 [Feature]: Experimental native RX 580 / gfx803 support for Qwen3.6 on Linux

Descrição

### What do you want to do?

Run Qwen3.6-35B-A3B through Strata on an AMD Radeon RX 580 (8 GiB, native gfx803) with system-RAM expert offload on Linux, without architecture spoofing. An explicit experimental, build-from-source option would make this older hardware usable without implying that every Polaris card or model is qualified.

The current workaround is a project-local port using isolated ROCm 6.3.4 and a separately built gfx803 rocBLAS/Tensile library. Draft PR #2765 proposes the narrow enablement changes: build selection, compatibility/kernel fallback, bounded uploads from pageable memory, and conservative hybrid prompt batching. Q8 KV, independent CPU speedups and startup caching are intentionally separate.

On the tested RX580 / 32 GiB RAM system, the bounded path passed a real API tool round-trip, stream cancellation/recovery and process restart. Streamed expert/MMQ prompt batches faulted and remain excluded; historical model-logit differences are unresolved. The draft documents these limits and does not claim general numerical or long-context qualification.

This issue tracks the requested experimental hardware support and discussion of its acceptance requirements. The draft still needs affected-backend build checks and evidence for the repository's unchanged-default-path requirement before maintainer review/merge readiness. No automatic setup, driver installation or broad supported-hardware claim is requested in this first change.